Remix.run Logo
▲ nlsvb 19 minutes ago

I have trouble taking seriously any claims that LLMs are conscious in any reasonable sense. I’ll try to make a few arguments quite succinctly.

1. There is zero dynamic nature to anything happening within the purview of the LLM. All dynamic elements come from external dynamic systems. E.g. every input produces the same mathematical output. We dynamically sample the output in order to produce dynamic results for the same input. For agentic use cases, the dynamic nature comes from tool calls to dynamic systems. All outputs for any set of model weights could be theoretically computed with pen and paper, so unless you were to claim such a process would be conscious in some sense, then emergent consciousness would have to come from some interaction between the substrate and the computation. But…

2. For efficiency, forward inference passes are done in batches. I.e. 100 different next-word predictions are being performed by the network at a time. For a single request, its batch position will not even be stable between each autoregressive step. The GPU is on 100 phone lines at once and each line switches conversations every step. Worse, with MOE, the attention and feed-forward parts aren’t even happening on the same GPU, even within the same step. There is no stable substrate here whatsoever.

3. There’s obviously the whole Chinese Room argument. The token embeddings being processed by LLMs are arbitrary high-dimensional vectors. The geometric relationships between these vectors are learned by the model, but the meaning is only ascribed once the output is decoded back into human language for grounded human interpretation, which all happens outside the model. In their current state, no argument about models developing world models based on this geometric structure makes any sense. There are infinitely many inputs / worlds that would produce isomorphic geometry.

▲consensus1 4 minutes ago | parent [-]

The vectors have meaning whether or not you can understand it